Skip to content
GenAI Learn/Capstone: Designing Full GenAI Systems
Browsing as a guest. Sign in to save your progress and earn XP as you complete chapters.

Worked Example: Search & Tutoring Systems

8 min read

You'll learn to

  • -Walk a full AI search engine design end to end
  • -Walk a full AI tutor design end to end
  • -See a real mastery-based adaptive teaching decision, computed live

A Perplexity-style AI search engine needs to search the web, synthesize an answer with real citations, and do it for a million queries a day with genuine real-time freshness. This is RAG, but with two twists Tier 8 did not have to deal with directly: the corpus is the entire live web, not a fixed document set, and citations are not a nice-to-have, they are the entire product.

AI Search: Citations Are Not Optional Decoration

The synthesis prompt has to force citations structurally, "answer using only information from the provided sources, and cite [Source N] after every claim", the exact grounded-generation pattern from Tier 8's citation tracker chapter. And exactly as in that chapter, generating a citation is not the same as generating a correct one: an output guard has to verify every single [Source N] reference actually maps to a real retrieved source before the answer ever reaches the user.

Freshness Changes the Caching Calculus

Tier 10's caching chapter treated a 24-hour TTL as a reasonable default. A search engine promising real-time freshness cannot use that default. A cache TTL of an hour or less is the tradeoff this specific product actually needs, accepting a meaningfully lower cache hit rate, and therefore higher cost, in exchange for answers that do not go stale on fast-moving news. This is exactly the kind of requirement-specific tuning step 1 of the framework exists to surface before a design gets built around the wrong default.

AI Tutor: Adaptive Teaching, Not a Static FAQ Bot

500,000 students across 20 subjects, each learning at a genuinely different pace, is precisely the personalization problem Tier 9's long-term memory chapter was built for, applied here to tracking what a specific student does and does not yet understand, rather than what a specific user prefers for dinner.

A real mastery-based teaching decision, computed for three different students

The Socratic method requirement is a real prompting constraint, not just pedagogical flavor text. The system must never simply give the answer, it has to ask a question that leads the student to discover it themselves, "what happens when you multiply both sides by three?" rather than "the answer is x equals nine." This has to be enforced directly in the system prompt, since a general-purpose instruction-following model defaults to being maximally helpful in the most direct way possible, which for a tutor is often exactly the wrong behavior.

Paranoid Safety, Same as Every Child-Facing System This Course Has Built

This is the same paranoid-sensitivity guardrail configuration from Tier 7's safety chapter and Tier 10's production guardrails chapter, applied here without exception, since the users are children. Every response passes through age-appropriate content checking before it ever reaches a student, with no lower-stakes exception carved out anywhere in this particular system.

Interview Signal is part of Pro

See a real weak answer next to a real strong one for this exact topic.

Quiz is part of Pro

Test what you just read with a short quiz, and bank the XP.

Ready to Build This?

Build the AI Search Engine and AI Tutor levels in the GenAI Lab.

ScaleDojo Logo
Initializing ScaleDojo